AI can already identify your writing style as a fingerprint, someone used a hairdryer on a weather sensor to game Polymarket for $34k, and London has been paying rent with horseshoes since 1211. [DVYio Newsletter](https://newsletter.dvy.io/visit/oXDiCTh8) Tech [Polymarket chart showing the 22°C temperature prediction spiking from nearly 0% to 100% in a vertical line on April 15.](https://newsletter.dvy.io/visit/Nj25v5UR) [Someone used a hairdryer on a Paris airport weather sensor to win $34,000 on Polymarket](https://newsletter.dvy.io/visit/Nj25v5UR) Someone appears to have walked up to a weather sensor at Charles de Gaulle airport with a hairdryer, briefly heated the air around it, and pocketed $34,000 from Polymarket bets as a result. French police are now investigating. Polymarket ran daily markets on peak Paris temperatures, settling on readings from a single Météo France sensor near the CDG runway. On April 6th and April 15th, that sensor recorded late-evening spikes well above the day's otherwise steady readings (21°C and 22°C respectively, against an 18°C baseline that had held all day). Someone had bought the long-shot temperature outcomes cheaply beforehand, since nobody expected them to hit. The sensor spiked, the markets resolved, they cashed out. Météo France noticed physical changes to the instrument, ran the data, and filed a complaint with the Air Transport Gendarmerie Brigade. The investigation is ongoing. The structural problem here isn't the hairdryer. It's that the entire market rested on a single unmonitored sensor in a publicly accessible location. Once you know that, the attack is almost obvious. Polymarket presumably didn't think anyone would bother, and for most of its history that was probably a reasonable assumption. Prediction markets tend to attract people who are committed enough to find this kind of edge. [A shadowy giant with a glowing infinity symbol over its face looms over a worker, absorbing glowing data from their hands.](https://newsletter.dvy.io/visit/M9kPiXJ3) [Meta is training AI on how its own employees use computers](https://newsletter.dvy.io/visit/M9kPiXJ3) Meta is rolling out a tool called the Model Capability Initiative that records how employees use their work computers: every keystroke, every mouse click, all of it logged and fed into AI training data. The idea is that if you want AI to help people use computers, you need examples of real people actually using them. Meta says the data is only used for that purpose, and that sensitive content is filtered out. The timing is making people uncomfortable. Meta has already cut around 2,000 jobs this year, its job listings have collapsed from roughly 800 in March to seven now, and employees are bracing for more. Having your smallest actions harvested for the AI that may eventually replace your role does rather sharpen the stakes. One current employee called it "very dystopian". For context: apparently Meta already had the ability to access employee activity on company machines. What's new is the explicit, systematic logging of that activity specifically to train AI models. Meta plans to spend around $140 billion on AI in 2026, roughly double last year's figure. The company is clearly treating this as an infrastructure problem: to build capable AI agents, you need training data, and the most convenient source of realistic computer-use data is the people already sitting at the computers. [Surreal digital sculpture of organic textures: pink petal-like forms, dripping amber liquid, and a branch of small fruit.](https://newsletter.dvy.io/visit/3QzgrNg2) [The sense AI researchers forgot to build: smell](https://newsletter.dvy.io/visit/3QzgrNg2) Your nose knows roughly 1 trillion distinct smells. Your eyes can distinguish maybe 10 million colours. Smell isn't a quirky bonus sense; it's one of the most information-dense channels humans have, tied directly to memory, emotion, and even early disease detection (some people lose their sense of smell years before a Parkinson's diagnosis). So why isn't anyone building it into AI? Turns out: almost nobody is trying. Research into artificial olfaction (giving machines a sense of smell) has barely moved in a decade, while vision and language research has exploded. The big AI conferences largely ignore it. The headlines saying "AI can now smell" are doing a lot of heavy lifting: what those systems actually do is repeat patterns from text, like knowing that humans describe blue cheese as pungent, without detecting a single molecule. A hospital AI that could detect infection from breath samples, or a food safety system that could flag spoilage before a label says so, would need genuine olfactory ability. Right now, that doesn't exist, and few people seem bothered by the absence. [A light blue grid background with white words arranged in the shape of a circular fingerprint.](https://newsletter.dvy.io/visit/eXnzUmWA) [Your writing style is a fingerprint AI can already read](https://newsletter.dvy.io/visit/eXnzUmWA) Most writing advice focuses on finding your voice. Turns out your voice can also find you. Journalist Megan McArdle ran a simple test after seeing someone online report that Claude Opus 4.7 had identified them as the author of unpublished fiction from a thousand-word sample. McArdle fed the model the opening chapter of a romance novel she'd written nearly twenty years ago, during what she describes as the hysterical phase of a bad breakup. Claude named her correctly. She then kept trimming the sample to find the floor. At 1,441 words, it still got there. The model wasn't doing anything mystical. It picked up on specific stylistic patterns: the sentence rhythms, the wry ironic framing, the "free indirect discourse" (the narrative trick where a character's thoughts blend into the narrator's voice without quote marks), and social preoccupations that matched her published nonfiction. It didn't need to search the internet. It recognised the fingerprint already in its training data. The implication for anyone who writes pseudonymously, or posts fiction to stay separate from their professional identity, is uncomfortable. Your unpublished drafts may be safer than your published ones, but only because the model hasn't seen them yet. Once you've published enough under your own name, the pattern is already there. This Week I Learnt [Someone holds a polished billhook and axe above a table containing horseshoes and a pile of iron nails.](https://newsletter.dvy.io/visit/SA5JTWN6) [Since 1211, London has paid rent on a lost plot of land in Shropshire with a billhook and an axe](https://newsletter.dvy.io/visit/SA5JTWN6) [Every October, a representative of the City of London turns up at the Royal Courts of Justice and hands over a billhook, an axe, six horseshoes, and 61 nails. Not as a quirky tourist stunt. As rent, legally owed on a lease signed in 1211. The payment goes to the Queen's Remembrancer, the oldest judicial post in England, a role created specifically to track what was owed to the Crown. The Remembrancer accepts the tools, confirms the debt is settled for another year, and that's it. The ceremony has run more or less the same way for eight centuries. The odd part: nobody knows where the land actually is. The original parcel was somewhere in Shropshire, but the records haven't survived, and the location has been lost for centuries. London is still faithfully paying rent on a property it cannot find, to a landlord that no longer really exists, in tools instead of money. The lease just... never ended.](https://newsletter.dvy.io/visit/SA5JTWN6) Poll On balance, has social media been good for the world? Positive Negative Pretty neutral Too early to tell Last Week's Poll Would you prefer the newsletter to be longer, shorter, or is it the right length? Make it longer 25% Keep it as it is 69% Make it shorter 6% History [A 1927 Paramount map shows California and Nevada repurposed as global film sets, with the Nile River, Swiss Alps, and Venice, Italy, mapped onto the West Coast.](https://newsletter.dvy.io/visit/7ymnBEUW) [Paramount Filming Locations (1927)](https://newsletter.dvy.io/visit/7ymnBEUW) [In 1927, Paramount published a map of California and Nevada listing locations that could double as exotic international settings, the idea being to keep productions on home soil and avoid the cost of sending a crew abroad. A California coastline could become a Welsh fishing village. A Nevada desert could become North Africa. On paper, at least. Peter Atwood mapped the original and noticed a few things. Some locations are plausibly chosen. Others suggest nobody at Paramount headquarters had spent much time in Kentucky or New England. It's also unclear whether the list was aspirational (here's what we could use) or retrospective (here's what we already have). Pre-sound era films are hard to track down, and a quick dig didn't settle it. The black-and-white point is notable. Strip out colour and you lose a lot of the cues we use to place a landscape. Underexpose a foggy California morning and it genuinely could pass for somewhere bleaker. Atmosphere and contrast did a lot of heavy lifting that colour film would have immediately undermined.](https://newsletter.dvy.io/visit/7ymnBEUW) [Title page for The Encyclopaedia Britannica, Eleventh Edition, Volume 1, published by Cambridge University Press in 1910.](https://newsletter.dvy.io/visit/MBKmz8iG) [The 1911 Encyclopædia Britannica is now fully searchable online](https://newsletter.dvy.io/visit/MBKmz8iG) The 1910–1911 Britannica is often called the greatest edition ever published. Not because it was the most accurate (it wasn't, by modern standards), but because it was written by actual experts at the peak of their fields: people like T.E. Lawrence on crusader castles, Algernon Swinburne on poetry, and Peter Kropotkin on anarchism. The writing has a confidence and depth that later editions quietly dropped in favour of being more accessible. What makes it useful today isn't the science (treat that with caution) but the historical perspective. It captures how educated, Western-centric minds understood the world on the eve of the First World War, before two world wars, decolonisation, and a century of revisionism reshaped everything. Reading an entry on, say, Africa or the Ottoman Empire tells you as much about 1910 as it does about the subject itself. This version is fully searchable and cross-referenced, which makes it far more usable than a physical set. Worth keeping bookmarked if you do historical research or just want to understand how a particular idea was framed before the modern consensus formed around it. Science [Aerial view of a narrow strip of coastal land and sparse homes surrounded by encroaching ocean and marshland.](https://newsletter.dvy.io/visit/YHecwz2v) [Southern Louisiana is going underwater](https://newsletter.dvy.io/visit/YHecwz2v) Not metaphorically, not eventually: the shoreline is projected to move 30 miles inland of New Orleans, and the entire coast will become uninhabitable. The question isn't whether people leave, it's whether they leave in an orderly way or get displaced in a panic. A new study in Nature Sustainability, led by researchers at Yale and Tulane, argues Louisiana has a shrinking window to do this properly. "Managed retreat" is the term: a multigenerational, planned movement of people, homes, and infrastructure to higher ground, rather than the chaotic scramble that tends to follow a disaster. Think of it less like evacuation and more like a slow rewrite of where a region exists. The researchers point to past examples in Louisiana itself: communities that moved after Katrina, and Indigenous groups in the Mississippi Delta who've been adapting to shifting land for generations. The Biloxi-Chitimacha-Choctaw Tribe on Isle de Jean Charles is one concrete case, relocated inland through a federally funded initiative after the island lost 98% of its land since the 1950s. The core argument is about timing. Act now, while there's still a functioning community to plan around, and you can design affordable housing, sensible infrastructure, and real support for people on the move. Wait for the crisis, and you get displacement: rushed, expensive, and brutal on the people least able to absorb it. Louisiana, for all the bleakness of its situation, has a chance to build the playbook before everyone else needs it. [A diagram comparing a nanorobot's four-lobed flower shape and its microscopic image to a four-rotor drone.](https://newsletter.dvy.io/visit/zWwuCMd8) [Robots smaller than a bacterium that herd bacteria with light](https://newsletter.dvy.io/visit/zWwuCMd8) Shine a laser at these nanorobots and the recoil from individual photons nudges them forward, a bit like how a garden hose pushes back against your hands when water shoots out. Twist the polarisation of the light (the angle at which the light waves oscillate) and the robot turns. No wires, no chemical fuel, no contact. The robots are less than one micrometre across, roughly 50 times smaller than the width of a human hair, and they carry tiny antenna structures that absorb and redirect incoming light to generate thrust. At that scale, the mass is so low that even photon recoil produces real speed and acceleration. The practical target right now is bacteria. Researchers can use these robots to locate, capture, and reposition individual bacteria in liquid, which sounds straightforward until you consider that doing this with conventional tools is a bit like trying to pick up a grain of sand using a crane the size of a building. Getting fine control over individual microbes opens up possibilities in drug-resistance research, targeted therapies, and the kind of single-cell biology that currently relies on much blunter instruments. [A magnifying glass held over a spreadsheet reveals three identical columns of data, indicating copy-pasted results.](https://newsletter.dvy.io/visit/GmWjXwY7) [How much dodgy science is hiding in plain sight in public datasets?](https://newsletter.dvy.io/visit/GmWjXwY7) Researchers scanning the raw Excel files behind 600 published papers found 18 datasets with serious copy-paste errors. One belongs to a landmark 2016 Parkinson's study that's been cited over 3,000 times and helped establish the idea that the disease originates in the gut. The underlying data contains sequences of duplicated values that should belong to completely different individual mice. It had been sitting in an open-access repository for eight years before anyone flagged it. The detection method: a researcher built a program to catch exactly this kind of duplication, partly inspired by two high-profile data fabrication cases where copy-pasted blocks were hiding in plain sight. The software isn't doing anything exotic. It's looking for patterns a careful human reviewer could spot, if they'd thought to look. The peer review system assumes someone checks the data. In practice, reviewers mostly read the paper, not the spreadsheet. A study can accumulate thousands of citations while the numbers underpinning it sit unchallenged in a public repository. If one tool scanning 600 datasets found 18 serious concerns, the question isn't whether this problem is widespread. It's how widespread it is. [A turquoise lap pool sits under a coffered ceiling with warm, glowing light panels that reflect off the water's surface.](https://newsletter.dvy.io/visit/ja2notnJ) [A long, dark indoor swimming pool framed by narrow stone pillars, with a coffered ceiling and warm lighting.](https://newsletter.dvy.io/visit/ja2notnJ) [A white cushioned lounge chair set back in a wood-panelled alcove featuring floor-to-ceiling cream curtains and a stone pool edge.](https://newsletter.dvy.io/visit/ja2notnJ) [Symmetrical marble double staircase with glowing step lights framing a dark corridor with intricate metal screens.](https://newsletter.dvy.io/visit/ja2notnJ) [A moodily lit indoor swimming pool with a coffered ceiling, private lounge alcoves, and reflections in the water.](https://newsletter.dvy.io/visit/ja2notnJ) [A symmetrical, dimly lit hallway with dark glass doors and warm orange lighting reflecting off polished marble floors.](https://newsletter.dvy.io/visit/ja2notnJ) [Hôtel Costes Spa](https://newsletter.dvy.io/visit/ja2notnJ) [Paris, France](https://newsletter.dvy.io/visit/ja2notnJ) Design [Three Claude Monet paintings above a color palette of ten swatches ranging from muted greys and tans to deep blues and greens.](https://newsletter.dvy.io/visit/XnvEgKFF) [Colour palettes pulled from 3,000 master paintings](https://newsletter.dvy.io/visit/XnvEgKFF) A searchable database of colour palettes extracted from over 3,000 paintings by the Old Masters. You can browse by artist, style, or genre (Renaissance, Expressionism, Orientalism, and so on), or use the colour harmony explorer to pick any hue and see which colours painters historically paired with it. [“Idea: An anonymous "vote to end meeting" button on Teams where if 50% of people press it, the meeting ends immediately.”](https://newsletter.dvy.io/visit/ywgYjLvC) [@OrevaZSN](https://newsletter.dvy.io/visit/ywgYjLvC) Data [Line chart showing the share of websites with AI-generated text rising from 0% before ChatGPT's launch in late 2022 to over 20% fully generated and 35% AI-assisted by April 2025.](https://newsletter.dvy.io/visit/GDBEFcTA) [35% of new websites since 2022 are AI-generated or AI-assisted](https://newsletter.dvy.io/visit/GDBEFcTA) [By mid-2025, roughly a third of newly published websites contained AI-generated or AI-assisted text, up from essentially zero before ChatGPT launched in late 2022. Researchers tracked new sites published between 2022 and mid-2025 using Internet Archive data to arrive at that figure. You might expect a flood of AI content to drag down factual accuracy or flatten everything into the same writing style, but the data doesn't support either of those. What it does show is a drop in semantic diversity (the range of ideas and meanings being expressed) and a nudge towards more positive-sounding language overall. One plausible explanation: most people spinning up AI-generated sites aren't trying to spread misinformation. They're either experimenting or filling pages to generate ad revenue. If you just want content that sounds plausible and fills space, a language model will oblige, and it defaults to producing mostly accurate, blandly upbeat text because that's what most of its training data looked like. Nobody's steering it towards lies, so it doesn't produce many. The worry, then, isn't so much that the web gets wronger. It's that it gets blander, and that the range of human perspectives and ways of expressing ideas quietly narrows.](https://newsletter.dvy.io/visit/GDBEFcTA) Random [A dark-themed video editor interface with a large central preview player, source import panels, and a multitrack timeline at the bottom.](https://newsletter.dvy.io/visit/3KcHRLkE) [VidStudio: browser-based video editor that never uploads your files](https://newsletter.dvy.io/visit/3KcHRLkE) A free, browser-based video editor that processes everything locally on your device. Nothing gets uploaded to a server, which means your files stay private and you're not waiting on slow transfers. It's powered by WebAssembly (a technology that lets browsers run near-native-speed code, normally the domain of installed apps), so it's genuinely fast rather than the sluggish experience you might expect from an in-browser tool. Resize footage for YouTube, TikTok, or Instagram; trim clips; batch-convert files to MP4; compress to hit specific file size limits; extract or mix audio; generate thumbnails and storyboards; add watermarks with positioning and motion controls; burn in subtitles; and edit on a multi-track timeline with frame-accurate seeking. No account, no install, works in any modern browser. [An aerial view of a massive warehouse construction site with tilt-up concrete walls surrounding a dirt foundation.](https://newsletter.dvy.io/visit/nc3ygmM8) [How America's warehouses became the perfect tool for disappearing people](https://newsletter.dvy.io/visit/nc3ygmM8) Warehouses are everywhere (and that's partly the point). They cluster at the edge of every mid-sized city, beside every major freight corridor, built to process goods fast and leave no particular impression on the people driving past. Logistics companies track every pallet and parcel that moves through their loading bays, but they'd rather you didn't track them. That calculated invisibility has a new consequence. The U.S. Department of Homeland Security is repurposing some of these buildings as makeshift detention centres, holding people without charge before deportation. The same design features that made them useful for next-day delivery (remote location, massive footprint, minimal public-facing presence) now make them useful for something much darker. Charmaine Chua's argument is that warehouses have always been instruments of control, not just storage. Understanding who owns them, who zones the land they sit on, and who profits from their expansion is, practically speaking, the first step to understanding how to slow that machine down. [A large white ship crane lowering a yellow tracked ROV into the ocean for subsea cable repair.](https://newsletter.dvy.io/visit/HSq2Ruhr) [Fixing a broken cable at the bottom of the ocean](https://newsletter.dvy.io/visit/HSq2Ruhr) There are two kinds of subsea cable worth knowing about. Data cables are essentially bundles of glass fibres, each about the width of a human hair, wrapped in layers of protection. The closer to shore, the bulkier the casing; deep-sea sections are thinner and buried further down. Signal boosters (called repeaters) sit along the route every 40–80 km to stop the light pulses fading. Power cables are chunkier, ranging from 70 mm to over 210 mm in diameter, and come in AC or DC flavours. AC works fine for runs up to about 80 km; anything longer and you need DC at higher voltage, because AC loses too much energy over distance. The most common cause of damage isn't sharks or earthquakes (though both have happened), it's ship anchors and fishing trawlers dragging gear across the seabed. When a data cable breaks, engineers send light pulses down the fibre and measure how long it takes to bounce back from the break, which pinpoints the fault location. Power cables use electrical fault-detection equipment instead, then send down an ROV (a remotely operated underwater vehicle) to find the exact spot. From there, a specialist repair ship heads out, which can take days or weeks depending on where the fault is and what the weather is doing. The crew lifts the cable, cuts out the damaged section, splices in a replacement, and lowers it back down. The whole process is slow, expensive, and deeply unglamorous. Given that these cables carry the vast majority of international internet traffic, it is also quietly critical infrastructure. [A TGV train interior with rounded blue and terracotta booths, a red floor, and a fluted glass railing by a walkway.](https://newsletter.dvy.io/visit/CwzJd6MR) [A TGV train interior with blue leather bench seating, featuring rounded backrests and a small wooden café table.](https://newsletter.dvy.io/visit/CwzJd6MR) [A yellow, mushroom-shaped table lamp with a curved head is mounted to a train window ledge above a wooden table.](https://newsletter.dvy.io/visit/CwzJd6MR) [A modern TGV train interior with minimalist blue and grey upholstered seats and wooden tray tables.](https://newsletter.dvy.io/visit/CwzJd6MR) [TGV Redesign](https://newsletter.dvy.io/visit/CwzJd6MR) [by Nendo](https://newsletter.dvy.io/visit/CwzJd6MR) [Web](https://newsletter.dvy.io/visit/2AGoBz8J) · [Twitter](https://newsletter.dvy.io/visit/89qftheW) · [LinkedIn](https://newsletter.dvy.io/visit/H73a9TfQ)